Automation is the reason to evaluate it
FaceFusion documents separate commands for interactive use, headless execution, batches and queued jobs. That gives a technical evaluator a concrete question: can its command and job workflow reproduce the media processing their team needs?
The README lists headless-run, batch-run and commands to create, submit, run, retry and delete jobs. Those are documented capabilities, not workloads verified by our sandbox. We did not measure output quality, processing speed, webcam behavior or accelerator compatibility.
Follow the documented installation path
The README warns that source installation needs technical skills and is not recommended for beginners. It also points readers who are uncomfortable with a terminal to Windows and macOS installers. We did not evaluate those installers.
The installation guide initializes Conda and creates an environment with Python 3.12 and pip 25.0. It then covers platform preparation, accelerator setup, cloning the repository, installing the application and reactivating the environment. The supplied guide lists CPU, CUDA, CoreML, DirectML, MIGraphX and OpenVINO installation choices. It ends with python facefusion.py run --open-browser.
Treat those instructions as the next step for a trial. Our retained harness output does not show which installation command ran or establish that the documented application setup completed.
What our recorded run establishes
Our stored run began on September 30, 2026 at 12:30:19.677 UTC and used commit 72470819a0373be3388b3929c8f8f311f418fc3c, abbreviated 7247081. The release page associates 3.9.1 with that abbreviated commit. The recorded sandbox had 3 CPUs, 8 GB of memory and image ghcr.io/astral-sh/uv:python3.12-bookworm. The install tail identifies Python 3.12.12 at /work/home/venv.
The harness marked install successful in 29 seconds and build successful in 6 seconds. It recorded 87 dependencies and a 780 MB installation metric. The checkout metrics were 247 files, approximately 23,048 lines of source and 2.1 MB. The checkout detector found one CI workflow and no Dockerfile or Compose file. That observation does not establish whether separate container distribution options exist.
The install tail contains only the Python environment line, and the build tail is empty. Without the executed commands or fuller logs, neither successful step proves that FaceFusion launched, an installed binary worked, model downloads completed or a media job succeeded.
The test step exited 1 after 46 seconds of harness time. The retained pytest summary says 13 failed, 97 passed, 4 skipped and 128 errors in 39.83 seconds. The harness duration and pytest's reported duration describe different timing scopes.
The parser also records total=238. Its counts have not been reconciled against a complete log to establish exhaustive unique coverage. The 128 errors should therefore remain a reported error count, rather than become 128 additional failed tests or a verified count of collection errors.
The retained tail shows NoneType path errors in face-tracking, frame-storage, job, video and image tests; five inference-pool assertion failures; and a missing test module in the static-provider test. It does not establish their root causes. A failed test command in this environment does not prove the application is broken on a supported installation.
Pip-audit recorded 58 advisories associated with 3 packages. The record does not provide scanned-package coverage, package names, advisory identifiers, severity or evidence that vulnerable code was reachable. These are recorded audit findings requiring investigation, not proof of exploitation or 58 independently verified product vulnerabilities. Dependency count is not audit coverage.
Check model terms before choosing a preset
The license page says FaceFusion is distributed under OpenRAIL-AS and that supplied assets have individual licenses. Its model list includes MIT and Apache 2.0 entries alongside Non-Commercial, ResearchRAIL, S-Lab 1.0 and Unknown entries. Application licensing alone cannot settle whether your selected model combination is suitable for commercial work.
Record the exact models and revisions you intend to use, then review their terms for your intended use. The project's disclaimer also says the software is designed to abstain from processing nudity, graphic content and sensitive material. That is a maintainer statement about safeguards; our run did not test their effectiveness. Use consented material and establish permission to use the depicted people before evaluating output.
Two alternatives for different workflows
Deep-Live-Cam's README documents image/video processing and a webcam sequence: select a source face, choose a camera and start live mode. It is relevant when live swapping is the main task. Its setup and performance claims are not comparative measurements from this review.
Rope's README describes a GUI implementing InsightFace's inswapper_128 model, with settings for folders and models. It is relevant when that GUI and model workflow match your needs. The extract does not support calling Rope a leaner alternative.
How to read the scores
These are editorial evidence ratings, not benchmark results or failure probabilities. The scale is 1 for insufficient evidence to assess a category, 2 for substantial documented prerequisites or unresolved validation, 3 for a documented baseline, 4 for broad documentation or corroborated operational evidence, and 5 for broad evidence plus representative validation. An unknown category receives 1 because this review cannot substantiate a stronger assessment, not because the product has been shown to perform poorly.
Setup receives 2 because the documented path requires environment and accelerator decisions while our retained output leaves application setup unverified. Documentation receives 3 because the supplied README and installation, licensing and disclaimer pages cover a baseline evaluation path; we have not verified every linked page. Community receives 1 because the supplied extracts do not establish support responsiveness, contributor breadth or adoption. Maturity receives 2 because release 3.9.1 is documented, but the recorded test and audit results remain unresolved and representative media processing was not measured.
For a useful next trial, follow the official installation instructions on your intended hardware, record the commands and model choices, and process a small consented sample. Check the output and reproduce relevant tests before deciding whether the workflow meets your requirements.

